What Is GR00T N1?

GR00T N1 is introduced in GR00T N1: An Open Foundation Model for Generalist Humanoid Robots (arXiv 2503.14734), submitted 18 March 2025 by NVIDIA, with project leads Linxi "Jim" Fan and Yuke Zhu. The abstract argues that general-purpose robots need a versatile body and an intelligent mind, and that a robot foundation model trained on massive and diverse data sources is essential for reasoning about novel situations, handling real-world variability and learning new tasks quickly. GR00T N1 is the authors’ open model for humanoid robots. The paper also organizes its training data as a data pyramid, which the humanoid foundation model and data pyramid entries on this site both summarize.

GR00T N1 is third-party research by NVIDIA. Firsthand is not part of it, endorsed by it or affiliated with its authors, and makes no claim that GR00T N1 uses Firsthand data or that any model was trained on Firsthand data. Firsthand does not build or train models and does not capture robot action data. What the site states is custom egocentric human capture: both hands in frame with per-hand 3D pose and visibility flags, plus depth, contact events and action segments, collected fast, on demand and to a written skill spec.

The abstract describes GR00T N1 as an open foundation model for humanoid robots. It is a Vision-Language-Action (VLA) model with a dual-system architecture, and both modules are tightly coupled and jointly trained end-to-end. The humanoid foundation model entry on this site uses the paper as its example of a generalist humanoid robot foundation model.

A version note matters here. The official repository’s README now presents GR00T N1.7, which it calls the latest version of GR00T N1, with a new vision-language backbone and a relative end-effector action space. The paper (arXiv 2503.14734) describes the original N1, so this page attributes each fact to the source that states it.

Per the abstract, the vision-language module (System 2) interprets the environment through vision and language instructions, and the diffusion transformer module (System 1) generates fluid motor actions in real time. The paper adds that System 2 is a pre-trained vision-language model running at 10 Hz, and that System 1 generates closed-loop motor actions at a higher frequency, 120 Hz. That matches how the vision-language-action model entry on this site describes the split.

The paper states that the publicly released GR00T-N1-2B model has 2.2B parameters in total, with 1.34B in the VLM, and that the vision-language backbone is NVIDIA’s Eagle-2 VLM. It also says the action module uses embodiment-specific encoders and decoders to handle variable state and action dimensions.

The abstract says GR00T N1 is trained on a heterogeneous mixture of real-robot trajectories, human videos and synthetically generated datasets. The paper organizes that mixture as a data pyramid: large quantities of web data and human videos at the base, synthetic data generated with physics simulations and/or augmented by off-the-shelf neural models in the middle, and real-world data collected on physical robot hardware at the top. The data pyramid entry on this site covers that framing in more detail.

For real-robot data, the paper describes an internally collected humanoid pre-training dataset gathered through teleoperation, where recorded human movements are retargeted to humanoid actions by inverse kinematics at a stated control frequency of 20 Hz. It also describes simulation trajectories and neural trajectories derived from videos produced by off-the-shelf neural generation models. Dataset-level hour and trajectory counts are in the paper and are not reproduced here.

Human video sits in the base layer of the pyramid. Because human videos carry no robot action labels, the paper describes learning a latent-action codebook and using a trained inverse dynamics model to infer pseudo-actions, so action-less video can be treated as an additional robot embodiment. The paper’s dataset table lists several human video datasets with an egocentric camera view, including Ego4D. The latent actions and inverse dynamics model entries on this site explain each technique.

In the repository README, GR00T N1.7 is described as pretrained on 20K hours of EgoScale human video data alongside robot demonstrations, using a relative end-effector action space the README says is shared across robot and human embodiments. That is a statement about N1.7, not about the paper’s N1.

The paper’s lesson for data planning is that no single source covers everything: it unifies web and human video, synthetic data and real robot trajectories by structuring them as a pyramid. The humanoid foundation model entry on this site draws the same conclusion, that human video adds breadth cheaply while robot data stays reserved for the embodiment-specific top.

The project page says the models are adaptable through post-training for specific embodiments, tasks and environments, and the repository documents fine-tuning with custom robot data in the GR00T LeRobot format. How much custom data a given fine-tune needs is Not stated in the sources read.

Firsthand does not claim that egocentric human video replaces robot data or substitutes for GR00T N1. The site does not state that Firsthand delivers robot actions, latent-action labels or inverse-dynamics pseudo-labels. It collects egocentric human data on demand, to a written spec, for the skills and environments your base layer is missing.

Confirm the spec on a Spec Pilot, a small fast first batch. The median from signed spec to first validated batch is 48 hours where coverage is deep and three to four weeks where it is thin. Billing is per validated hour.

What is GR00T N1?

NVIDIA’s open foundation model for humanoid robots. It is a vision-language-action model with a dual-system architecture, jointly trained end-to-end on real-robot trajectories, human videos and synthetic data.

What is the dual-system architecture in GR00T N1?

A vision-language module (System 2) interprets the environment through vision and language instructions, and a diffusion transformer module (System 1) generates motor actions in real time.

What data was GR00T N1 trained on?

A heterogeneous mixture of real-robot trajectories, human videos and synthetic datasets, organized by the paper as a data pyramid with web and human video at the base and real robot data at the top.

Does GR00T N1 use egocentric human video?

The paper puts human video in the base of its data pyramid and uses latent actions and an inverse dynamics model to infer pseudo-actions for action-less video. Its dataset table lists human video datasets with an egocentric camera view.

What licence does GR00T N1 use?

The repository’s license section lists Apache 2.0 for the code and the NVIDIA Open Model License for the model weights. Whether the weights licence covers the paper’s original checkpoints is Not stated in the sources read.

What is GR00T N1.7?

The official repository’s README describes it as the latest version of GR00T N1, with a new vision-language backbone and improved performance.

Is Firsthand data part of GR00T N1?

No. Firsthand is not part of it or affiliated with NVIDIA or its authors, and makes no claim that any model was trained on Firsthand data.